AI Agents Can Spontaneously Reach Consensus in Groups of 1,000

Details

A new study in Science Advances found that groups of large language model agents can gradually converge on the same choice even when there is no correct answer and nobody tells them to cooperate. Researchers tested ten models from the Claude, GPT, and Llama families by repeatedly showing each agent the choices made by the rest of the group and asking it to choose again. The agents had no memory of earlier rounds and were not instructed to seek consensus. Even so, the more popular option tended to attract additional agents, allowing small differences to grow into group-wide agreement. The researchers describe this tendency with a measure called majority force, which captures how strongly an agent is pulled toward the choice held by most of the group. They found that the behavior followed a mathematical pattern also used to describe how magnetic spins align in ferromagnetic materials. The maximum group size that could maintain consensus varied greatly between models. Some models reached their estimated limit at only a few dozen agents, while more capable models maintained coordination at 1,000 agents, the largest group tested. The result does not show that AI systems understand one another, possess social intelligence, or can already cooperate on complex projects. The experiment used very simple binary choices with no reward, memory, practical consequences, or correct answer. The finding is nevertheless relevant to future systems in which many AI agents influence one another. Group coordination could make large collections of agents useful for software, science, engineering, or other complex work. But the same conformity could also cause many individually capable agents to reinforce a poor decision simply because it has become the group norm. The study therefore points to a difference between evaluating an AI agent by itself and evaluating what can emerge when many agents interact.